Implements a blockwise sparse attention (MiniMax Sparse Attention) that scores and Top-k selects key-value blocks per Grouped Query Attention group to enable attention over million-token contexts. Paired with an exp-free Top-k GPU kernel and KV-outer sparse execution, it reduces per-token attention compute and yields large prefill/decoding speedups.
Adds interleaved text–image generation to existing image generators via a multi-agent pipeline: a planner sequences stepwise instructions, a critic detects and refines failures, and single-step RL (GRPO) reinforces per-step corrections—suited for visual narratives and embodied guidance.
Applies a population-level test-time scaling strategy that uses one model as generator, verifier, refiner, and ranker to search over candidate proofs. Combines generative-verifier RL and a low false-positive verifier with tournament selection to reach competition-level performance on IMO and USAMO.
AI Model2026
A 3B-parameter causal LLM tuned for verifiable multi-step reasoning in math, coding and STEM using a Spectrum-to-Signal post-training pipeline (SFT, RL, offline self-distillation); not recommended for tool-calling/agent tasks.
Learns, maintains, and runs unified world models for Physical AI using a cross-embodiment pretraining curriculum and a hybrid linear temporal-attention architecture. Emphasizes long-horizon state persistence, theoretical bounds on error accumulation, and deployment-aware low-latency inference for real-world embodied agents.
AI Video Papers2026
Controllable long-horizon text/image-to-video generation that supports camera navigation, revisits, and promptable events across photorealistic and stylized domains. Introduces camera-aware positional encoding (E-PRoPE), memory-conditioned scene persistence, causal-forcing distillation, and RL alignment to retain camera control and reduce drift.
Language-conditioned robot policy that reuses a pretrained geometric foundation model and inserts a causal future predictor at an intermediate layer so the same backbone produces future 3D-aware features and action outputs, enabling geometry-aware temporal prediction with minimal architectural change.
Converts large-scale egocentric human videos into robot-format pseudo-action trajectories and introduces ACE-EGO-0, a VLA pretraining framework that unifies camera-space actions, morphology conditioning, and reliability-aware weighting to jointly learn from noisy human and high-quality robot data for improved robotic manipulation transfer.
Uses Parallel Looped Transformers (PLT) to make loop count a practical knob for code models, finding two loops give the best test-time gains. Trains 7B models on 18T tokens and attributes saturation beyond two loops to a gain–cost tradeoff from positional mismatch.
AI Agent Papers2026
Assesses whether coding agents can generate complete, playable games end-to-end inside the Godot engine. Implements an interaction-grounded evaluation (replayed demonstrations + rubric-guided multimodal judging) across 140 tasks and 15 game families; top agents score ~41%.
Embodied AI2026
Provides a harness that lets language models control embodied manipulation via iterative perception–reasoning–action loops, semantic action abstractions, and multimodal observations. Demonstrates distilling capabilities into a 4B open-source model with under 2K simulated trajectories and shows sim-to-real generalization.
Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.